How can I remove blur from a picture that shook

Camera shake is a smear with a direction. Remove the blur from the picture and check a straight edge that should have stayed straight.

The search I typed that night was how can i remove blur from a picture, lowercase, from the kitchen, with the camera still in the bag and the tripod still in the other bag. I had stood on a pavement in the early evening and shot a green-tiled shopfront at 1/15 of a second. The frame looked fine on the camera back. On the laptop it had a vertical smear, the same smear in the tiles and in the coat of a person who was not walking.

TL;DR. The blur was camera shake, vertical, about 22 pixels on a 6240 by 4160 frame, because I shot at 1/15 with a 35mm lens and left the tripod at home. Horizontal grout doubled. Vertical door edges stayed comparatively single. That pattern is not the same as a cyclist moving through a sharp street, which is the next frame on the card. I converted the raw file with sharpening off, tried a short vertical deblur, and kept the pass where the grout became one line and the shoulder did not grow a ghost. The camera JPEG was the wrong place to start, because it had already sharpened the smear.

The frame, the lens, and the bag

The camera was a Fujifilm X-T3. The lens was the 35mm f/2, which on that body is a short portrait field, about 50mm in old 35mm terms. The file I cared about was frame 1912. The raw file is street-mercer-1912.RAF, 43 megabytes. The in-camera JPEG is street-mercer-1912.jpg, 6240 by 4160, 7.8 megabytes, sRGB, 1/15 second, f/2, ISO 1600.

I knew the shutter was slow while I was standing there. The tiles had just caught the last warm light, and the interior of the shop was already a dark green. If I had gone to 1/60 I would have been at ISO 6400 or I would have opened a lens that was already open. I chose the smear over the noise, and I told myself I was braced. I was leaning on the rail of a bus stop. Bracing is not a tripod. The tripod was a travel Gitzo, folded, in the flat, because I had decided this walk was only a walk. The picture happened anyway.

The old handheld guide for this lens, on this sensor, is a shutter around 1/50 or faster if I want edges to stay put. I was at 1/15. That is more than a stop and a half slower than the guide, and I had walked there with a coffee. I am not surprised by the file. I am annoyed that I checked the camera back at arm's length, where a 22 pixel streak does not show, and called it sharp.

There is a person in the frame, standing, dark coat, paper cup, face turned toward the shop. They were not the reason I shot. The tiles were. They matter anyway, because a repair that redraws a stranger's face is a failure even if the grout looks crisp. I treat the coat and the face as a test, not as a portrait I was hired to flattering. I was not hired at all. It is still a person.

What 100 percent showed

At thumbnail size the JPEG looks like an evening street picture with a little softness, the kind you blame on f/2. At 100 percent it is not that. Focus was close enough. The smear has a direction.

I looked at five places, in the same order, and wrote the notes into street-mercer-1912.txt so I would not relitigate them the next morning.

The chrome door handle is a small highlight that should be a blob a few pixels across. It is a vertical streak. I read the coordinates off the raw conversion, not the JPEG. Top of the bright core at y = 1840, bottom at y = 1862, x confined to a band about 3 pixels wide. Twenty-two pixels, vertical. That streak is the cleanest measure of the shake I have in the frame, because a highlight is as close as the street gets to a point.

The pale grout between the green tiles runs horizontally. Each line is doubled: two dark edges and a lighter core, the pair separated by roughly 8 to 10 pixels. I do not treat that gap as the kernel length. Grout has its own width, and a doubled edge of a thick line understates the travel of a point. The handle is the ruler. The grout is the confirmation that the travel was vertical. Edges that run across the motion smear. Edges that run with the motion mostly stay thin.

The door frame is a dark vertical. It is still one line. It is a little longer at the ends than it ought to be, and a little soft, but it is not a pair of door frames side by side. If the shake had been horizontal I would have seen the opposite: the door doubled, the grout still single. I did not see that.

The neon in the window is a red tube, meant to be a thin horizontal stroke with a short vertical return at one end. The horizontal stroke is a fat band. The vertical return is still skinny. Same direction test, second sample, so I am not betting the whole repair on one door handle.

The coat is noisy and the edge of the shoulder is doubled in the same vertical way as the grout. The face is soft along the jaw, vertically, and the eyes are not two different smears from the tiles. If the person had turned their head during the exposure, the jaw would be worse than the tile. It is not. The tile is just as bad. The camera moved.

The handle crop I measured twice

I cropped the handle out of the raw conversion as street-mercer-1912-handle-crop.tif, 400 by 600 pixels, 16-bit, 1.4 megabytes. I measured the streak on the full frame, then again on the crop, because I have misread coordinates on a dark file before. Both passes gave 22 pixels, give or take a pixel at the faint tail of the streak. The tail is dimmer than the core. If I include the dim tail I can talk myself into 26. I used 22 as the working length and I tried neighbors rather than pretending the tail was a precise instrument.

That crop is also what I show anyone who looks at the camera JPEG and says the picture is fine. The JPEG, displayed small, hides it. The crop, displayed at 100 percent, is a bright dash that should have been a dot.

The next frame is a different blur

Frame 1913 is street-mercer-1913.RAF, same lens, 1/125 second, f/2, ISO 1600. I had stepped back and the light was the same, so this one is noisier in the shadows and it is not shaken. The grout is one line. The door handle is a blob. The shop is usable.

A cyclist crossed the corner during that 1/125. The rider is a horizontal stain, on the order of 70 pixels, wheels stretched along the road, no face to recover. The background does not share that stain. This is subject motion. The shutter was fast enough for my hands and far too slow for the bike.

How can i remove blur from a picture is a question that hides this split. A single slider named blur does not know which of the two frames you held up. On 1912, a vertical kernel can be a fair model of the whole frame, because the whole frame moved together. On 1913, a horizontal kernel long enough to fit the cyclist will drag the sharp shop into echoes. A kernel fit to the shop will leave the cyclist as a cyclist-shaped smear, which is the correct failure. There is no single place the rider was. The exposure added many places together.

I did not run a deblur on 1913. I cropped the shop as street-mercer-1913-shop.jpg, 4200 by 2800, 5.1 megabytes, and left the rider out of that crop. If I ever want the rider I will print the smear and call it a smear. I will not ask a tool to invent a face that was never in one place.

The standing person on 1912 is the control for this lesson. They were a subject, and they were sharp relative to their own motion, which is to say they were as blurred as the door and no more. "There is a person in the picture" does not mean the blur is subject motion. The still edges have to agree.

Why I refused the camera JPEG as the starting file

The in-camera JPEG is the file I would have grabbed, because it was 7.8 megabytes and it was right there. At 100 percent the grout already had a bright rim and a dark rim. The camera had sharpened the JPEG in its usual way. Sharpening a smear does not pull the smear back into a point. It outlines whatever edges survived, including the doubled grout and the long handle.

I converted street-mercer-1912.RAF in a raw processor with sharpening at zero, noise reduction at zero, and a plain white balance from the tiles. The export was street-mercer-1912-base.tif, 16-bit, 6240 by 4160, 156 megabytes. That size is the check. 6240 times 4160 times six bytes is about 156 megabytes. When the export landed at 156, I believed it was 16-bit and uncompressed. A 50 megabyte TIFF would have been an 8-bit save, and I would have started over.

The base TIFF looks duller than the JPEG and it looks less "crisp." It is the honest file. The smear is smooth instead of outlined. A deblur has a chance of treating the handle as a streak of light rather than as a drawn line. I kept the JPEG in the folder as camera-jpeg/ so I could remember what not to repair. I did not delete it. Clients and future-me both ask for the JPEG first.

ISO 1600 on this camera, with noise reduction off, is grain in the coat and in the shop interior. That grain is part of the base file on purpose. A deblur boosts whatever is mixed into the smear, and noise is mixed in. I would rather see the worms and decide, than let the camera smooth the coat into plastic before I have even tried.

Three kernel lengths, and the one I kept

I am not going to pretend I measured a perfect kernel and the picture snapped back. I tried three lengths on copies of the base TIFF, all of them vertical, all of them aimed at the shake and not at a general softness.

street-mercer-1912-k18.tif used a short vertical run of 18 pixels. The handle got shorter and was still a streak. The grout was still a double line if I looked straight at it, though a thumbnail looked improved. I rejected it for the grout, not for the thumbnail.

street-mercer-1912-k22.tif used 22. The handle collapsed toward a bright mark with a faint tail. The grout read as one line at 100 percent. The neon tube thinned back toward a tube. The standing person's shoulder stayed one shoulder. The eyes did not move, and they did not gain a catchlight the shop window did not justify. The coat noise went slightly wormy, vertical, in the darkest fold, and nowhere else that I could find in a ten-minute look. This is the file I kept as the repair.

street-mercer-1912-k30.tif used 30, which was me including the dim tail and then adding a little more out of pessimism. The grout grew a dark echo above the line, a classic ring. The shoulder grew a second contour offset upward. The neon developed a pale duplicate. The coat worms became obvious enough to see without zooming all the way in. A longer kernel was not a stronger repair. It was the wrong kernel, and the wrong kernel is louder than the blur.

I did the tries through the unblur tool on a JPEG exported from the base TIFF, street-mercer-1912-base.jpg, quality 95, 11.2 megabytes, because I did not need to upload 156 megabytes to learn the length. Once 22 was the one I believed, I repeated that pass and saved the result next to the base TIFF. The tool file is a prediction with a direction in mind, not a laboratory deconvolution with a kernel I typed in by hand. I still judged it as if the length mattered, because the three results behaved the way wrong and right lengths behave. The notes on classical kernels versus a learned pass live in how to deblur a picture. This street frame only needed the practical split: 18 short, 22 plausible, 30 ringing.

I did not follow the repair with a global sharpen. The k22 file was already near the edge of worms in the coat. How to sharpen a blurry picture is a later, milder decision, and on this frame I skipped it. The outlined JPEG from the camera is what sharpening first looks like. I did not need a second copy of that mistake.

What the print size does to the same 22 pixels

A streak of 22 pixels feels disastrous at 100 percent on a 6240-wide file, because 100 percent makes every pixel a screen pixel and the file is much larger than the screen. On a print it is a length, and the length depends on how wide I print.

At 12 inches wide, 22 pixels is 12 times 22 divided by 6240, which is about 0.042 inches, a millimeter. At a normal viewing distance that is softness, not an obvious dash. At 6 inches wide it is half of that, and I would have struggled to defend the repair to anyone but myself. The smear is real at every size. It is not equally worth a risky pass at every size.

The ask, when it came, was not a 12 inch print of the whole frame. It was a crop of the neon and the tiles. I made street-mercer-1912-neon-crop.tif from the base file first, 2200 by 1600, so I could see the damage at the size of the ask. The 22 pixel streak is the same 22 pixels, now inside a 2200-wide picture instead of a 6240-wide picture. Printed at 8 inches, that streak is 8 times 22 divided by 2200, about 0.08 inches. Still small, and now plain to see, because the crop is mostly flat tile and a thin tube. Flat areas and thin bright lines are where a short shake shows. A busy hedge would have hidden it.

I made the same crop from k22. The tube is one tube. The tile lines are single. The dark window above the neon is a little wormy if I press my face to the print, and fine at arm's length. I delivered the k22 crop as street-mercer-1912-neon-crop-k22.jpg, 2200 by 1600, 2.4 megabytes. I did not deliver the full frame repair as a replacement for the raw file. The RAF stays the negative. The base TIFF stays the unsharpened conversion. The k22 file stays a derivative with a note that says "vertical, 22, coat slightly wormy, face unchanged."

Before I trust the direction, the frame has to pass four plain checks:

• A highlight that should be a point is a streak, and I can measure its length.
• Edges that cross the streak are doubled, and edges that run with it stay single.
• The standing person is no worse than the tiles.
• A faster frame of the same shop, if I have one, does not share the streak.

File	What it is	Size	What I did with it
street-mercer-1912.RAF	Camera raw	43 MB	Kept, never overwritten
street-mercer-1912.jpg	Camera JPEG, already sharpened	7.8 MB	Kept as a warning, not a master
street-mercer-1912-base.tif	16-bit conversion, no sharpen	156 MB	The file I actually repaired
street-mercer-1912-k18.tif	Vertical try, too short	156 MB	Rejected, grout still double
street-mercer-1912-k22.tif	Vertical try, handle matches	156 MB	Kept as the repair
street-mercer-1912-k30.tif	Vertical try, too long	156 MB	Rejected, rings and a ghost shoulder
street-mercer-1913-shop.jpg	Next frame, shake-free crop	5.1 MB	Used for the shop, cyclist excluded

The failure I almost kept

On the first k22 pass I looked at the neon, nodded, and nearly exported. I had not looked at the coat fold. The worms there are the same height as the kernel, which is the tell. Real cloth texture does not line up with the shake direction at a regular pitch. The worms do. I saw them on the second pass, in the morning, which is why the morning pass exists.

I made one local attempt on a duplicate, street-mercer-1912-k22-coat.tif. I smoothed only the darkest fold, with a soft mask, and stopped when a button lost its rim. The button is four pixels of highlight. It is not important to the tile picture. It is important as a stop sign. If a cleanup eats a highlight I can name, it will eat one I cannot name. The delivered crop does not include the coat, so the coat cleanup never needed to be finished. I still kept the note. The next shake file might be a portrait, and the coat lesson is the portrait lesson.

I also checked the standing person's eyes at 200 percent on k22 and on the base TIFF, toggling. Same shape, same dull catchlight from the street, no new iris texture. A repair that "improves" eyes on a 22 pixel shake of a face that was already small is inventing a person. This face stayed the soft, small, real face. Good.

The page I use when the blur is general and I do not have a direction is fix blurry photo. I mention it because I almost opened that mindset on 1912 and treated the frame as uniformly soft. It is not uniformly soft. It is soft across horizontal edges and much less soft along vertical ones. A general soft-focus pass would have fattened the door frame that did not need it, and it would have done less for the grout than a vertical pass did. Direction is the whole diagnosis. How to fix blurry photos collects the wider set of causes. This note is only the vertical shake, with the tripod in the bag and the proof in the handle.

What I would do on the pavement next time

I would look at the camera back and then zoom, or I would shoot a second frame at 1/60 and ISO 6400 as insurance. A noisy sharp negative of these tiles would have beaten a clean smear, because noise is easier to live with than a doubled grout when the subject is a flat wall of line. I did not shoot that insurance frame. Frame 1913 is insurance for the shop by accident, at 1/125, and it is not the same composition. The warm skim of light on the tiles is only on 1912. That is why I repaired 1912 instead of pretending 1913 replaced it.

I would also put the tripod in the bag I actually carry, or I would admit I am not going to, and I would raise the shutter. Both are honest. "I will hold still" at 1/15, on a rail, after a walk, was not honest. The file is the record of that.

If the handle streak had been 80 pixels instead of 22, I would not have kept a repaired version. The mixture would be too long, the rings would show at the print size of the crop, and the tiles would be a guess. I have a separate note on why a long smear has no single true scene to return to. For 22 pixels, on a regular tiled wall, with a highlight to measure, the repair was a fair attempt and the k22 crop is the one I will print. The raw file remains in the folder, unmodified, because the attempt is not the negative.